9559e7d4 - Build the error-fix fixer: drive an error.fix conclusion to a draft PR unattended

DESIGN.md

@@ -419,6 +419,7 @@ agent watch grok-usage                         # one scan; knock child (under th
 agent watch assigned [--follow]                # allowlisted GitHub assignments → runner session + knock
 agent watch errors                             # one scan; $AGENT_HOME/error-fix.json; knock daemon polls with grok-usage
 agent watch error-fix                         # one scan; find-or-create implement task + isolated worktree; knock daemon polls with grok-usage
+agent watch error-fix-work                    # one scan; drains error-fix implement tasks from spec_written through PR gates to done (§21.7); not wired into agent daemon
 agent supervise --session ID [--repo OWNER/REPO --number N] [--once|--follow]
 agent status
 agent dashboard [--port 7845]
@@ -497,7 +498,7 @@ The rules below were already implied by §§1–17. They are now explicit so a l
 
 A worker report such as “analysis complete” or “tests passed” is **input**. It is not the transition. Opening a draft is not done.
 
-No transition that needs deterministic evidence may be satisfied by model text alone. Malformed structured output is rejected (unknown `activity.type` → `execution_status=error`; empty, partial, timeout, or unavailable review output is not zero findings). A patch that does not apply is a failed check, not a debate.
+No transition that needs deterministic evidence may be satisfied by model text alone. Malformed structured output is rejected (unknown `activity.type` → `execution_status=error`; empty, partial, timeout, or unavailable review output is not zero findings). A patch that does not apply is a failed check, not a debate. The structured `FINDINGS:` section is the sole source of truth for a review verdict; a real finding stated only in a reviewer's free-text preamble and not restated there is not mechanically detectable and is accepted residual risk, not something a heuristic tries to close.
 
 ### 19.2 Untrusted inputs
 
@@ -665,10 +666,26 @@ The model never receives production credentials. Analysis that only reads the ex
 
 `agent watch error-fix` find-or-creates the implement task and clones `https://github.com/<repo>.git` into `$AGENT_HOME/error-fix-work/<task_id>`; `agent github pending` still opens drafts, and a retry draft uses the existing head `error-fix-<id8>`.
 
+`spec_written` stays human-only for ordinary implement tasks (`HUMAN_KEYS`, step kind `human`). Exception: when the task is error-fix-originated — payload carries `error_id` **and** a matching validated `error.fix` activity exists for that `error_id` in the same session — `close-step --source script` may set `spec_written=ja` with evidence. Evidence is still mandatory. An `error.seen` or `error.skip` alone does not qualify.
+
+For confirmed error-fix tasks only (same `error_fix_confirmed` condition), `close-step --source script --status n_a` may close `deviation_declared` / `deviation_granted` with evidence. Any other status (e.g. `--status ja`), or any task without a confirmed `error.fix` origin, stays human-only. Without this carve-out no error-fix task can reach `done`, since both keys are `HUMAN_KEYS` with no other script path to close them.
+
 ### 21.6 Not in this revision
 
 - A second hub state machine, leases, or autonomous merge
 
+### 21.7 Automated fixer driver
+
+`agent watch error-fix-work` drains open error-fix `implement` tasks on this device (`payload.error_id` set and a matching `error.fix` confirmed for that id in the same session, state not `done`/`failed`) from `spec_written` through a draft `pr.open`, the PR gates (`grok_pr_quality`, `grok_pr_logic`, `codex_pr_quality`, `codex_pr_logic`, plus the scripted `contributing_ok` carve-out), and task state `done`, using only script control flow and the `grok`/`codex` CLIs via `lane.launch()`. A human still merges the PR; the spine has no `merged` step. It is not wired into `agent daemon`.
+
+- Scripts the five-part spec under `$AGENT_HOME/error-fix-specs/<task_id>/.spec.md` (a sibling of `error-fix-work`, never inside the pushed git worktree) from the `error.fix` brief plus `error.seen` metadata (never raw log excerpts), closes `spec_written` via the script carve-out above, then `agent round start`.
+- Walks the spine with the same step executor as `agent run` (including auto pass/fail for reviewer and PR-reviewer lanes from `STATUS:` + `FINDINGS:`). Round retries reset the relevant checklist keys to `nein` and call `agent round start`. Cap is `task.current_round` against 5: exceeding it sets `task state failed` and stops touching that task.
+- When both dimensions of one vendor PR-review pair (`grok_pr_quality`/`grok_pr_logic`, or `codex_pr_quality`/`codex_pr_logic`) are ready simultaneously, the driver prepares both on the store-owning thread, launches any still-pending dimensions concurrently via a thread pool (worker threads only call the lane launch itself, never touch the store), then fully finishes both on that same thread (gate row, checklist, agent finish — no abandon/discard). Task-level continue-vs-message and the combined rejection-feedback write happen once afterward, aggregated across the batch so either dimension's rejection is preserved regardless of pair order; a `failed` outcome in the batch skips the deferred `round start` and wins over a sibling's `cont=True`. On an unhandled exception mid-prepare/launch/finish, any still-working agent row for either dimension is released before the exception propagates, and a rejection reset already committed earlier in the batch still receives its deferred `round start` (best-effort) when no outcome failed the task. `agent round start` itself refuses `state=failed` the same way it refuses `state=done`. `agent gate record --verdict rejected` also leaves `state=failed` unchanged (rather than its usual auto-transition to `implementing`) when the sibling already failed the task in the same batch — the rejected gate row is still recorded either way, for audit, even though the task stays permanently stopped. Ordinary `agent run` and every other spine step remain one-at-a-time.
+- If a vendor CLI binary is missing (`OSError` / `FileNotFoundError` before any `LaneResult`) or a lane returns `LaneResult(status="unavailable")` on both the initial attempt and the one retry, the driver leaves task and checklist state untouched for retry, but releases any already-started agent row (`cmd_agent finish --verdict unavailable`) rather than leaving it `working` forever — notes the CLI looks unavailable, and moves on; the next scan retries after a human fixes PATH/auth.
+- Each scan re-checks from the ledger (not per-call local state) whether `pushed` is closed but no successful (`done`) `pr.open` activity row exists yet for that task's branch head. A mid-flight `pending` row is resumed via `scan_github` (no duplicate insert); an `error` row or missing row triggers a fresh `insert_pr_open_and_scan` — so a failed insert is not silently skipped by the next scan.
+- After `pushed`, inserts a pending `pr.open` (title/body per CONTRIBUTING) and runs `agent github pending`, then continues through the PR gates to `done`.
+- Failing a task via lane retry-exhaustion also finishes the still-working agent row (`blocked` for implementer, `rejected` for reviewer/pr-reviewer roles) so the row does not block a later manual round-start recovery.
+
 ## 22. Static supervise loop (v1)
 
 A second model must not orchestrate the first. `agent supervise` is a **script** with locked questions and locked answers. Model text is not a state transition.

README.md

@@ -77,7 +77,7 @@ agent pg status
 agent pg stop
 ```
 
-`agent run` records a local check when `local_check_pass` is open and the snapshot has no local checks yet (it does not rerun an existing failed check). It closes an agent step when the session store already has the artifact, and with `--spec-file` launches the vendor lane (tmux by default; `--no-tmux` for a subprocess). When `pushed` is open it git-pushes (no force) and closes with the HEAD sha; when `mergeable` is open it measures GitHub mergeability and checks and closes only if both are green. Reviewer lanes are not auto-approved from `STATUS: complete`.
+`agent run` records a local check when `local_check_pass` is open and there is no fresh pass/skip for the current HEAD (a prior fail on that HEAD is rerun). It closes an agent step when the session store already has the artifact, and with `--spec-file` launches the vendor lane (tmux by default; `--no-tmux` for a subprocess). When `pushed` is open it git-pushes (no force) and closes with the HEAD sha; when `mergeable` is open it measures GitHub mergeability and checks and closes only if both are green. Reviewer and PR-reviewer lanes auto-pass when `STATUS: complete` and `FINDINGS:` parses to zero.
 
 `agent github pending` is one scan: owned pending `pr.open`, `comment.post`, `review.post`, and `issue.write` rows via `gh`. Pull requests are drafts. A retry reuses an existing open draft, issue, or comment instead of creating a second one.
 
@@ -96,6 +96,7 @@ agent watch grok-usage  # one scan of SuperGrok weekly credits into usage.snapsh
 agent watch assigned [--follow]  # allowlisted assignments; needs `gh` and `$AGENT_HOME/watch.json`
 agent watch errors      # one scan; $AGENT_HOME/error-fix.json; no log host in this package
 agent watch error-fix   # one scan; find-or-create implement task + isolated worktree
+agent watch error-fix-work  # one scan; drains error-fix implement tasks from spec_written through PR gates to done; not wired into agent daemon
 agent supervise --session ID [--repo OWNER/REPO --number N] [--once|--follow]
 # agent knock (daemon, no --once) polls grok-usage, pending, pr.merged, github pending, mail pending, errors, and error-fix every 60s
 ```

src/agent_cli/chain.py

@@ -246,6 +246,25 @@ def required_source(step: Step) -> str:
     return "script"
 
 
+def is_error_fix_originated(snapshot: dict[str, Any] | None) -> bool:
+    """True when this is a validated error-fix task.
+
+    Requires both payload.error_id and snapshot.error_fix_confirmed (a matching
+    error.fix activity in the same session). Payload alone is not enough —
+    an error.seen / error.skip without error.fix must not get the
+    spec_written script carve-out.
+    """
+    from .error_fix_act import _nonempty_str
+
+    if not isinstance(snapshot, dict):
+        return False
+    payload = snapshot.get("payload")
+    raw = payload.get("error_id") if isinstance(payload, dict) else None
+    if not isinstance(payload, dict) or not _nonempty_str(raw):
+        return False
+    return bool(snapshot.get("error_fix_confirmed"))
+
+
 @dataclass(frozen=True)
 class CloseVerdict:
     allowed: bool
@@ -261,6 +280,7 @@ def close_allowed(
     source: str,
     evidence: str | None,
     snapshot: dict[str, Any] | None = None,
+    status: str = "ja",
 ) -> CloseVerdict:
     """May this key be set to ja/n_a NOW? Only the current next step."""
     step = find_step(workflow, key)
@@ -270,7 +290,18 @@ def close_allowed(
         return CloseVerdict(False, f"{key} requires evidence", step)
     want = required_source(step)
     if source != want:
-        return CloseVerdict(False, f"{key} requires --source {want} (got {source})", step)
+        # error-fix implement tasks may script-author spec_written (any status)
+        # and, when status=="n_a", deviation_declared/deviation_granted too —
+        # HUMAN_KEYS and Step.kind stay human so every other task keeps the
+        # human-only requirement for all three keys, unchanged.
+        script_carveout = source == "script" and is_error_fix_originated(snapshot)
+        deviation_na = (
+            key in ("deviation_declared", "deviation_granted") and status == "n_a"
+        )
+        if not (script_carveout and (key == "spec_written" or deviation_na)):
+            return CloseVerdict(
+                False, f"{key} requires --source {want} (got {source})", step
+            )
     ready = next_steps(workflow, checklist, spine_only=False)
     if step not in ready:
         pending = ",".join(s.key for s in next_steps(workflow, checklist, spine_only=True)) or "-"
@@ -300,10 +331,24 @@ def _latest_agent(snapshot: dict[str, Any], role: str, vendor: str | None) -> di
 def _latest_gate(
     snapshot: dict[str, Any], stage: str, dimension: str
 ) -> dict[str, Any] | None:
+    """Latest gate for stage/dimension.
+
+    `load_task_dict` orders gates by payload origin_seq (oldest→newest), so the
+    last matching entry is the true latest write. When snapshot.head_sha is set,
+    only gates for that exact head are considered — a stale approval for another
+    head must not satisfy the step (same strict scoping as local_check_pass).
+    When no head is bound yet, all gates for the stage/dimension are considered.
+    """
+    want = str(snapshot.get("head_sha") or "").strip().lower()
     hit = None
     for g in snapshot.get("gates") or []:
-        if g.get("stage") == stage and g.get("dimension") == dimension:
-            hit = g
+        if g.get("stage") != stage or g.get("dimension") != dimension:
+            continue
+        if want:
+            g_head = str(g.get("head_sha") or "").strip().lower()
+            if g_head != want:
+                continue
+        hit = g
     return hit
 
 
@@ -344,9 +389,40 @@ def _artifact_ok(step: Step, snapshot: dict[str, Any]) -> str:
         checks = list(snapshot.get("local_checks") or [])
         if not checks:
             return "no local_check recorded"
-        if any(c.get("result") == "fail" for c in checks):
+        want = str(snapshot.get("head_sha") or "").strip().lower()
+        if want:
+            # Scope to this head so a stale pass/fail for another (or empty) head
+            # cannot mask a fresh check — load_task_dict keeps the full history
+            # and same-second ran_at ties are otherwise order-unstable.
+            for_head = [
+                c
+                for c in checks
+                if str(c.get("head_sha") or "").strip().lower() == want
+            ]
+            # Last row per name wins (list is oldest→newest), same as unbound.
+            latest: dict[str, Any] = {}
+            for c in for_head:
+                name = c.get("name")
+                if name is not None:
+                    latest[str(name)] = c
+            latest_list = list(latest.values())
+            if any(c.get("result") == "fail" for c in latest_list):
+                return "local_check fail"
+            if not any(
+                str(c.get("result") or "") in ("pass", "skip") for c in latest_list
+            ):
+                return "no local_check for current head"
+            return ""
+        # No bound head: last row per name wins (list is oldest→newest).
+        latest: dict[str, Any] = {}
+        for c in checks:
+            name = c.get("name")
+            if name is not None:
+                latest[str(name)] = c
+        latest_list = list(latest.values())
+        if any(c.get("result") == "fail" for c in latest_list):
             return "local_check fail"
-        if not any(c.get("result") in ("pass", "skip") for c in checks):
+        if not any(c.get("result") in ("pass", "skip") for c in latest_list):
             return "local_check without pass/skip"
         return ""
     if step.key == "pushed":